Salesforce AI Pro Tips
↗️ Start with one domain: pick 3–5 high-volume intents (e.g., renewal nudges, refund checks, case summaries) before scaling.
↗️ Ground every agent in Data Cloud: map identities, unify consent flags, and use zero-copy to read warehouse data in real time.
↗️ Define actions, not just answers: expose Flow/MuleSoft APIs (refund, create case, schedule install) with clear inputs/outputs and limits.
↗️ Write tight system prompts: role, policies (do/don’t), escalation rules, and tone; store versions in a repo.
↗️ Guardrails first: permission sets for tools, field-level security, masked PII in logs, and spend limits per agent.
↗️ Human-in-the-loop: auto-escalate on low confidence, risky intents, or KBA failure; pass transcript + intent + proposed action.
↗️ Telemetry you’ll use: log intent, tools called, tokens, latency, success/fail, and business outcome (won deal, solved case).
↗️ Close the loop: pipe negative CSAT/escalations to QA (Watchtower/flows) and retrain prompts/actions weekly.
↗️ Automate with Flow before code: screens for guided ops, record-triggered for back office; reserve Apex for edge cases.
↗️ Use Slack as the cockpit: surface agent actions and record previews in deal/incident channels; notify on exceptions.
↗️ Analytics that matter: track Autonomous Resolution Rate, CSAT by intent, ARR influenced, time-to-first-response, and cost per resolution.
↗️ Path + Kanban adoption: enable Guidance for Success and stage-specific next steps to drive consistent agent outputs.
↗️ Data hygiene: dedupe leads/contacts, validate required fields, and enforce picklists so agents don’t write junk.